RAG & Search Engineering — Complete Reference

Build production-grade retrieval systems with modern hybrid search patterns.

This skill covers:

• RAG: Chunking, contextual retrieval, grounding, adaptive/self-correcting systems

• Search: BM25, vector search, hybrid fusion, ranking pipelines

• Evaluation: recall@k, nDCG, MRR, groundedness metrics

• Chunking strategies: Page-level chunking (0.648 accuracy, highest in NVIDIA benchmarks)

• Contextual Retrieval: Anthropic's 2024 technique (67% accuracy improvement with prompt caching)

• Hybrid retrieval: Lexical (BM25) + vector + cross-encoder reranking

• Reranking: Cross-encoder (ms-marco-TinyBERT-L-2-v2, 4.3M params, outperforms larger models)

• RAG evaluation: Recall@K, Precision@K, nDCG, groundedness, verbosity, instruction following

• Modern paradigm shift: Adaptive, multimodal, self-correcting systems (static RAG is over)

Key Insights:

• Page-level chunking achieved highest accuracy (0.648) with lowest variance
• Contextual Retrieval reduces retrieval failures by 67% when combined with reranking
• Semantic chunking improves recall by up to 9% over simpler methods
• Hybrid retrieval + reranking drastically improves accuracy
• Era of static RAG is over - adaptive, wise retrieval is mainstream

It focuses on doing, not explaining theory.

Scope note: Retrieval algorithm tuning (BM25/HNSW/hybrid, query rewriting) lives in ai-rag; this skill covers RAG-specific packaging, context injection, and grounded generation.

Quick Reference

Decision Tree: RAG Architecture Selection

Building RAG system: [Architecture Path]
    ├─ Document type?
    │   ├─ Page-structured? → Page-level chunking (0.648 accuracy, lowest variance)
    │   ├─ Technical docs? → Semantic chunking (9% recall improvement)
    │   └─ Simple content? → RecursiveCharacterTextSplitter (400-512, 85-90% recall)
    │
    ├─ Retrieval accuracy low?
    │   ├─ Multi-entity docs? → Contextual Retrieval (67% failure reduction)
    │   ├─ Noisy results? → Cross-encoder reranking (TinyBERT, <100ms)
    │   └─ Mixed queries? → Hybrid retrieval (BM25 + vector + reranking)
    │
    ├─ Dataset size?
    │   ├─ <100k chunks? → Flat index (exact search)
    │   ├─ 100k-10M? → HNSW (low latency)
    │   └─ >10M? → IVF/ScaNN/DiskANN (scalable)
    │
    └─ Production quality?
        └─ Full pipeline: Page-level + Contextual + Hybrid + Reranking → Optimal accuracy

When to Use This Skill

Claude should invoke this skill when the user asks:

• "Help me design a RAG pipeline."
• "How should I chunk this document?"
• "Optimize retrieval for my use case."
• "My RAG system is hallucinating — fix it."
• "Choose the right vector database / index type."
• "Create a RAG evaluation framework."
• "Debug why retrieval gives irrelevant results."

Related Skills

For adjacent topics, reference these skills:

• ai-llm - Prompting, fine-tuning, instruction datasets
• ai-llm - Agentic workflows, multi-agent systems, LLM orchestration
• ai-rag - BM25, hybrid search, ranking pipelines (complements RAG retrieval)
• ai-llm-inference - Serving performance, quantization, batching
• ai-mlops - Security, privacy, PII handling
• ai-mlops - Deployment, monitoring, data pipelines
• ai-prompt-engineering - Prompt patterns for RAG generation phase

Detailed Guides

Core RAG Architecture

• Pipeline Architecture - End-to-end RAG pipeline structure, ingestion, freshness, index hygiene, embedding selection
• Chunking Strategies - Modern benchmarks (page-level 0.648 accuracy, semantic, RecursiveCharacterTextSplitter 400-512)
• Index Selection Guide - Vector database configuration, HNSW/IVF/Flat selection, parameter tuning

Advanced Retrieval Techniques

• Retrieval Patterns - Dense retrieval, hybrid search, query preprocessing, reranking workflow, metadata filtering
• Contextual Retrieval Guide - Anthropic's 2024 technique (67% failure reduction), prompt caching, implementation
• Grounding Checklists - Context compression, hallucination control, citation patterns, answerability validation

Production & Evaluation

• RAG Evaluation Guide - Recall@K, nDCG, groundedness, RAGAS/TruLens, A/B testing, sliced evaluation
• Advanced RAG Patterns - Graph/multimodal RAG, online evaluation, telemetry, shadow/canary testing, adaptive retrieval
• RAG Troubleshooting - Failure mode triage, debugging irrelevant results, hallucination fixes

Existing Detailed Patterns

• Chunking Patterns - Technical implementation details for all chunking approaches
• Retrieval Patterns - Low-level retrieval implementation patterns

Templates

Chunking & Ingestion

• Basic Chunking
• Code Chunking
• Long Document Chunking

Embedding & Indexing

• Index Configuration
• Metadata Schema

Retrieval & Reranking

• Retrieval Pipeline
• Hybrid Search
• Reranking

Context Packaging & Grounding

• Context Packing
• Grounding

Evaluation

• RAG Evaluation
• RAG Test Set

Navigation

Resources

• resources/rag-evaluation-guide.md
• resources/rag-troubleshooting.md
• resources/contextual-retrieval-guide.md
• resources/pipeline-architecture.md
• resources/advanced-rag-patterns.md
• resources/chunking-strategies.md
• resources/grounding-checklists.md
• resources/index-selection-guide.md
• resources/retrieval-patterns.md
• resources/chunking-patterns.md

Templates

• templates/context/template-context-packing.md
• templates/context/template-grounding.md
• templates/chunking/template-basic-chunking.md
• templates/chunking/template-code-chunking.md
• templates/chunking/template-long-doc-chunking.md
• templates/retrieval/template-retrieval-pipeline.md
• templates/retrieval/template-hybrid-search.md
• templates/retrieval/template-reranking.md
• templates/eval/template-rag-eval.md
• templates/eval/template-rag-testset.jsonl
• templates/indexing/template-index-config.md
• templates/indexing/template-metadata-schema.md

Data

• data/sources.json — Curated external references

External Resources

See data/sources.json for:

• Embedding models (OpenAI, Cohere, Sentence Transformers, Voyage AI, Jina)
• Vector DBs (FAISS, Pinecone, Qdrant, Weaviate, Milvus, Chroma, pgvector, LanceDB)
• Hybrid search libraries (Elasticsearch, OpenSearch, Typesense, Meilisearch)
• Reranking models (Cohere Rerank, Jina Reranker, RankGPT, Flashrank)
• Evaluation frameworks (RAGAS, TruLens, DeepEval, BEIR)
• RAG frameworks (LlamaIndex, LangChain, Haystack, txtai)
• Advanced techniques (RAG Fusion, CRAG, Self-RAG, Contextual Retrieval)
• Production platforms (Vectara, AWS Kendra)

Use this skill whenever the user needs retrieval-augmented system design or debugging, not prompt work or deployment.